Volume 9, Issue 3, March – 2024
ISSN No:-2456-2165
Multi-Model Ensemble Approach for Soybean
Crop Yield Estimation (Kharif-2023) in
Latur District at Macroscale level
Ashutosh Pawar
Head, GIS and RS, Semantic Technologies and
Agritech Services Pvt. Ltd., Pune.
Priyanka Shamraj
Agrometeorologist, Semantic Technologies and
Agritech Services Pvt. Ltd., Pune.
Abstract:- Crop area estimation is a critical aspect of
agricultural monitoring and management, providing
essential information for decision-making in the
agricultural sector. Study was carried out at Semantic
Technologies and Agritech services Pvt. Ltd., GIS and
Remote sensing team, Pune during Kharif-2023. All
methodology given by YESTECH manual under Pradhan
Mantri Fasal Bima Yojana (PFMBY) was followed. Latur
district facing more weather-based yield losses during last
few of years. In this case study we tried to estimate yield
of soybean crop for agriculture-based stake holders,
insurance companies, Government polices at Revenue
circle level (RC). Multimodal approach is beneficial over
single model yield estimation approach as it takes
ensemble yield for perfect forecasting of crop yield.
Accuracy was in the range as given in YESTECH manual
at RC level. Thus, overall results show that use of such
model for yield estimation is one of the best approach to
take the decisions for insurance based stake holders in
rainfed regions where more negative consequences on
soybean productivity under different climate change
scenario was observed.
Keywords:- Remote Sensing, GIS, NPP, Machine Learning,
DSSAT-4.8, Soybean, Latur, Yield Simulation, Revenue
Circle, Soybean productivity.
I. INTRODUCTION
Agriculture is the backbone of global economies,
providing sustenance and livelihoods for billions of people.
The ability to accurately predict crop yield is paramount for
effective resource management, risk mitigation, and informed
decision-making. Traditional methods, reliant on historical
data and manual observations, often fall short in addressing
the dynamic nature of modern agricultural challenges. The
integration of advanced technologies has ushered in a new era
in agriculture, enabling a more nuanced and precise
understanding of crop dynamics. Software applications,
remote sensing, GIS, and AI/ML algorithms work
synergistically to process vast datasets, analyse patterns, and
predict crop yields with unprecedented accuracy.
IJISRT24MAR1981 www.ijisrt.com
International Journal of Innovative Science and Research Technology
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
Upasana Singh
Sr. GIS and RS Executive, Semantic Technologies and
Agritech Services Pvt. Ltd., Pune.
Bhargav Sonawane
Jr. RS Executive Semantic Technologies and
Agritech Services Pvt. Ltd., Pune.
Unpredictable rainfall, rising temperatures, and extreme
weather events like hailstorms and strong winds threaten crop
growth and yields. Farmers shift to drought-resistant crops,
rely heavily on irrigation, face soil degradation, and
experience economic vulnerability due to unstable
production.
In the contemporary agricultural landscape, the accurate
estimation of crop yield has emerged as a critical aspect
influencing various sectors including insurance, economy,
government policies, and ultimately, the welfare of farmers.
Traditional methods of crop yield estimation are often marred
by limitations in accuracy and efficiency. However, the
integration of advanced technologies, such as sophisticated
software, remote sensing, GIS (Geographic Information
System), and cutting-edge artificial intelligence (AI) and
machine learning (ML) techniques, has revolutionized the
precision and reliability of crop yield estimation.
 Importance of Accurate Crop Yield Estimation:
 Insurance Sector: Accurate crop yield estimates play a
pivotal role in the insurance sector, enabling precise risk
assessment and facilitating the development of tailored
insurance products.
 Economic Implications: Crop yield estimates are
fundamental to economic forecasting, impacting
commodity markets, trade agreements, and pricing
mechanisms.
 Government Policies: Governments rely on accurate crop
yield estimates to formulate effective agricultural policies.
This includes allocation of subsidies, distribution of
resources, and planning for strategic interventions during
periods of adverse weather conditions or pest outbreaks.
 Public Welfare and Food Security: Accurate crop yield
estimates are integral to ensuring food security and public
welfare.
 Farmers' Wellbeing: For farmers, precise crop yield
estimates translate into enhanced planning and risk
management. Access to reliable information empowers
farmers to make informed decisions regarding crop
2475
Volume 9, Issue 3, March – 2024
ISSN No:-2456-2165
selection, resource allocation, and market participation,
thereby improving overall farm productivity and
livelihoods.
The adoption of advanced methods for crop yield
estimation is a transformative step towards building
agricultural resilience in the face of evolving challenges. The
synergy between software applications, remote sensing, and
GIS technologies empowers stakeholders across sectors to
make informed decisions, fostering a sustainable and
prosperous future for agriculture. By recognizing the
multifaceted implications of accurate crop yield estimation,
societies can work collaboratively to strengthen the
foundations of global food security, economic stability, and
the welfare of farming communities.
This report delves into the significance of employing
advanced methods for estimating crop yield and highlights
their implications across diverse domains.
Fig 1: Study Area
IJISRT24MAR1981 www.ijisrt.com
International Journal of Innovative Science and Research Technology
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
II. MATERIAL AND METHODS
A. Study Area
Study was carried out at Semantic Technologies and
Agritech Services Pvt. Ltd., Pune during kharif season 2023
for particular assignment. For this study, all revenue circles
(RC) in the districts of Latur of Maharashtra state were used
as experimental sites. Field level data like ground truth, Crop
cutting experiments were carried out.
B. Geography and Climate for Latur District:
Latur districts cover 7,157 sq km of area. Annual rainfall
averages 520 mm, with Kharif season receiving 350-390 mm.
Kharif temperatures range from 33-37°C maximum and 22-
25°C minimum, with average relative humidity of 70-80%
located at latitudes 17°52′N to 18°50′N, 75°16′E to 76°42′E.
Elevations range from 400-800 m. Soil and Drainage can be
described as , Vertisols dominate the region, posing drainage
challenges due to flat topography and leading to waterlogging
during heavy rainfall. Situated in Boundary of Karnataka-
Maharashtra border, the district is surrounded by Osmanabad
in the South, Beed in the West, Parbhani and Nanded in the
North and Bidar district of Karnataka in the East. Soybean,
cotton, jowar, bajra, tur, and sugarcane is one of the major
crops which is taken in this district.
2476
Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

C. Methodology:  Semi Physical NPP- Net Primary Productivity

All methodology was followed by the procedure given  AI and Machine learning
by yield estimation system based on technology (yes-tech)  Crop simulation model-DSSAT-4.8
under Pradhan Mantri Fasal Bima Yojana (PMFBY).  Ensemble Model

Methodology used is multimodal approach for  Semi Physical Net Primary Productivity (NPP):
estimation of crop yield was given below. RC wise yield in
Tonnes/hector of soybean crop during kharif season 2023 was  Data and Materials Used:
estimated by all following methods. The data and materials used in this study are as follows:

Table 1: Data used for NPP Generation in Semi Physical Model

Data Satellite/Ground Resolution Source
Daily insolation/PAR INSAT-3D 4km resampled to 1km MOSDAC
10 days composite fAPAR ver. 2 PROBA V and SPOT-VGT 1km Copernicus Land
Service
8 Days Composite Surface Terra-MODIS 1km MODIS Time Series Tool
Reflectance
Paddy Mask Sentinel 1 5m USGS Explorer
Temperature Gridded data from 1km Interpolated NASA Power
NASA Power website
Light-use Efficiency Literature
Harvest Index Ground CCE

 Fraction of Absorbed PAR (FAPAR):

The FAPAR data is from Copernicus Land Service,
source link is (https: //land. copernicus. eu/global/index.
html). the 10 - day composite product with 1 km data is used.
The range of FAPAR lies between 0 and 1. The physical
values are retrieved from the Digital Number (DN).
 Photosynthetically Absorbed Radiation (PAR):
PAR is calculated from daily insolation data. The daily
insolation data is converted to 8 - day composite (sum) for the
whole period. 50% insolation is considered as PAR. This
daily insolation data is collected from MOSDAC from
INSAT - 3D satellite, source link (www.mosdac. gov. in) for
the crop season from 2018 to 2022.
Wstess = (𝟏−𝑳𝑺𝑾𝑰)/ (𝟏+𝑳𝑺𝑾𝑰𝒎𝒂𝒙)
The LSWImax value has been taken from the spatial
maximum of particular crop mask of the entire district.
 Temperature Stress:
Temperature Stress (Tstress): The daily average
temperature data is downloaded from NASA Power website,
source link is (https: //power. larc. nasa. gov/data - access -
viewer. html). It is a gridded data with a resolution of 1°0 *
1°0 latitude and longitude.

PAR= 8 - day composite * 0.5.

 Water Stress (Wstress):

The Wstress is calculated from Land Surface Water
Index (LSWI). The MODIS time series tool (MODIStsp) Where,
used to download and process the MODIS 8 day composite  Tmin = Minimum temperature required for the
(MOD09A1) source link is (https: //lpdaac. usgs. photosynthesis (°C).
gov/products/mod09a1v006), and LSWI is calculated for the  Tmax =Maximum temperature required for the
entire period with the formula . photosynthesis (°C).
 Topt = Optimal temperature required for the
LSWI = (ƿ𝑵𝑰𝑹−ƿ𝑺𝑾𝑰𝑹)/ (ƿ𝑵𝑰𝑹+ƿ𝑺𝑾𝑰𝑹) photosynthesis (°C);
 T = Daily mean temperature (°C).
LSWI value range from - 1 to 1, and higher positive
values indicate the vegetation and soil water stress. Further,
the Wstress is calculated from 8 days LSWI output –

Table 2: Data used for Soybean Crop for Semi-Physical Approach.

Sr.No. Particulars Values Source Sr.No. Particulars Values Source
1 T maximum 35°C (Nimje, P. 4 LUE 1.78 (Chavan et al., 2018)
2 T minimum 10°C M. 2022) 5 Harvest Index 0.45 Periodic CCE data.
3 T optimum 26°C

IJISRT24MAR1981 www.ijisrt.com 2477

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

On the off chance that air temperature falls beneath
Tmin, which is quite a rare chance than Tscalar value will
automatically become 0.
 Light Use Efficiency (ℇ):
The light use efficiency LUE is used for soybean crop
was 1.78 for the study.
 Crop Mask
The crop mask was derived utilizing Sentinel-1
synthetic aperture radar (SAR) data obtained from the
European Space Agency (ESA) Copernicus Hub. Employing
the R programming language, we employed the Random
Forest algorithm for the generation of the crop mask,
implementing hyperparameter tuning techniques and
contingency matrix analysis. This methodology was
systematically applied across our specified crops within the
targeted area of interest.
In terms of accuracy assessment, our results yielded a
robust accuracy range of 90% to 95% across all cultivated
crops and within various districts. This signifies a high level
of precision in delineating and classifying the specified crops
within the delineated geographical regions. The meticulous
incorporation of Random Forest algorithm, hyperparameter
tuning, and contingency matrix analysis has facilitated the
generation of a reliable and accurate crop mask, providing
valuable insights for agricultural monitoring and
management within the designated study area.
 Calculation of NPP and Grain Yield:
To compute the final Net Primary Productivity NPP and
its Grain Yield, the formula and equation is used as follows.
The NPP sum has been multiplied with Harvest Index (0.45)
to estimate per pixel yield.
NPP = PAR * FAPAR * ℇ * Tstress * Wstress (Logic of
Monteith Equation 1972).
Same methodology is followed by Upasana Singh et.al.
(2023) and also showing same results for all data used to run
the model.
 Crop Simulation Model-DSSAT
Crop simulation model is a mathematical equation or the
set of equations, which represents the behavior of system. We
used CROPGRO – for Soybean crop. It is consisting of various
subroutines viz., Water balance subroutine, Phenology
subroutine, Nitrogen subroutine, and Growth and
Development subroutine described below.
 Data Input to Model
Material and method and all file process was carried out
by the procedure followed by Hoogenboom, G., et.al (2019)
and (2024) Jones, J.W., (2003) and the minimum data
requirements for operation, calibration and validation of the
Crop models are described below.

Table 3: Showing List of Input Required by Crop Simulation Model

Sr.No. Input variables Acronym Source
1. SITE DATA
Latitude LAT NASA power
Longitude LONG NASA power
Elevation ELEV NASA power
2. DAILY WEATHER DATA
Maximum temperature TEMPMAX NASA power
Minimum temperature TEMPMIN NASA power
Solar radiation SOLARAD NASA power
Rainfall RAIN NASA power
3. SOIL CHARACTERISTICS
Soil texture SLTX
Soil local classification SLDESC
Soil depth SLDP
Colour, moist SCOM
Albedo (fraction) SALB
Photosynthesis factor (0 to 1 scale) SLPE
pH in buffer determination method SMPX DSSAT website
Potassium determination method SMKE Where Global gridded-
Horizon-wise soil profile dataset at 10-
Lower limit drained LL(L) km resolution was
Upper limit drained DUL(L) Developed for DSSAT-
4.8
Upper limit drained SAT(L)
Software crop simulation
Saturated hydraulic conductivity SWCN(L)
models.
Bulk density moist BD(L)
Organic carbon OC(L)
Clay (<0.002 mm) ` CLAY(L)
Silt (0.05 to 0.002 mm) SILT(L)

IJISRT24MAR1981 www.ijisrt.com 2478

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Coarse fraction (>2 mm) STONES(L)

Total nitrogen TOTN(L)
pH in buffer PHKCL(L)
Cation exchange capacity CEC(L)
Root growth factor 0 to 1 SHF(L)
4 MANAGEMENT DATA
Sowing date YRPLT Krishi-Dainandini
Plant population at seedling PLNATS Published by in Vasantrao
Planting method (TP/direct seeded) PLME Naik Marathwada Krishi
Row spacing ROWSPS Vidypeeth, Parbhani,
Row direction (degree from north) AZIR
Seed rate SDWTRL
Sowing depth SDEPTH
Irrigation dates IDLAPL
Irrigation amount AMT(J)
Method of irrigation IRRCOD
Fertilizer application dates FDAY(J)
Fertilizer amount N ANFER
Fertilizer type IFTYPE
Fertilizer application method FERCOD
Fertilizer incorporation depth DFERT
Tillage date TDATE
Tillage implements TIMPL

 Input files
The files are organized into input, output and
experiment performance data file. The experiment
performance files are needed only when simulated results are
to be compared with data recorded in a particular experiment.
In some cases, they could be used as input files to reset some
variable during the course of a simulation run. The input files
are further divided into those dealing with the experiment,
weather and soil and the characteristics of different
genotypes. Similarly output files are also further divided into
those dealing with the overview, summary, growth, water,
carbon and nitrogen balance.
 Soil properties directory file: The file SOIL.SOL contained
the list of different soils with their physical and chemical
properties.
 Soil profile initial condition file: The soil profile initial
condition file contained the initial values of soil water, soil
reaction and soil nitrogen data pertaining to this situation
was entered.
 Irrigation management file: The Irrigation management
file has the provision of date and amount per fixed
irrigation (mm) applied depth (cm) of management.
Irrigation data pertaining to this situation was entered.
 Fertilizer management file: The fertilizer management file
contained the date, form and amount of nitrogen
application. Accordingly, information on fertilizer
application was entered in the file.
 Treatment management file: The treatment management
file contained the description of each treatment under
separate title and serial numbers. The file also contained
dates of planting and emergence, plant population at
seeding and at emergence, planting method, planting
distribution, row spacing, row direction, planting depth,
planting material, transplant age, plants per hill, dates of
simulation beginning etc.
 Crop Cultivars Directory File
For Soybean CRGRO048 contained the list of different
cultivars with their genetic coefficients. The modified genetic
coefficients viz., CSDVAR, PPSEN, EMG-FLW, FLW-FSD,
FSD-PHM, WTPSD, SDPDVR, SDFDUR, PODDUR,
THRESH, SDPRO and SDLIP is used. Variety selected was
JS-335 which is mostly used in this area.
The genetic coefficients are the most important
parameters which represents the genetic characteristics of the
cultivar and on which the crop phenology, biomass
production partitioning and yield potential of the crop
depends. However, the actual performance is controlled by
the external factors also.
 Running the crop model: Once, all the desired files were
created carefully the model was run for all the crops
cultivars. Each run of model created output files.
 Machine Learning:
Methodology and processing of model is described
below in details.
 Data Collection and Ground Truthing:
 Collect remote sensing data (optical and radar imagery)
for the study area, covering the growing season of the
crops.
 Ground truth data collection using field surveys using
CropTech App ( prepared by compony) for accurate
calibration and validation.

IJISRT24MAR1981 www.ijisrt.com 2479

Volume 9, Issue 3, March – 2024
ISSN No:-2456-2165
 Crop Mask Extraction:
 Pre-process the remote sensing data to correct for
atmospheric interference and geometric distortions.
 Apply image enhancement techniques to improve the
visual quality of the images.
 Employ supervised or unsupervised classification
algorithms to extract crop masks for Soybean fields.
 Generation of Spectral Indices and use of RADAR
Backscatter:
 Calculate vegetation indices (e.g., NDVI, NDRE,
GNDVI) from the optical remote sensing data to assess
crop health and Vigor.
 Utilize backscatter data from radar imagery to analyse
surface roughness and other relevant crop information
(VV, VH).
Fig 2: Methodology used in Machine learning Approach
 Ensemble Models
This methodology aims to combine the predictive power
of both Machine Learning (ML) models and Crop Simulation
Models (CSM) to provide an enhanced and more accurate
estimation of crop yields. Here is a structured approach:
IJISRT24MAR1981 www.ijisrt.com
International Journal of Innovative Science and Research Technology
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
 Crop Cutting Experiments:
 Use of Crop Cutting Experiment (CCE) for Crop with
smart sampling methods to efficiently estimate crop
parameters for crop.
 Training and Testing Models (Machine Learning):
 Divide the dataset into training and testing sets, ensuring
no overlap between the two.
 Evaluate the model's performance on the testing dataset
using evaluation metrics like accuracy, F1-score, and
mean squared error (RMSE).
 Model Validation and Final Result:
 Validate the trained model using independent ground truth
data collected during the growing season for Soybean.
 Assess the model's accuracy and generalization ability to
ensure reliable yield estimation.
 Obtain the final crop yield estimation results for Soybean
in the study area.
 Data Collection and Preprocessing:
 Gather data from both ML, Semi-Physical Approach and
CSM approaches as outlined in the above methods.
 Consolidate all input data: weather data, soil properties,
crop management practices, spectral indices, RADAR
backscatter, and ground truth data.
 Ensure data alignment in terms of temporal and spatial
granularity.
2480
Volume 9, Issue 3, March – 2024
ISSN No:-2456-2165
 Individual Model Generation:
 Machine Learning Approach:
 Utilize various algorithms like Linear regression, Random
Forest, Extra Trees, k-earest neighbours, and neural
networks.
 Train these models on the dataset ensuring proper
validation and calibration.
 Crop Simulation Approach
 Use well-calibrated crop simulation models such as
DSSAT.
 Simulate the growth and yield of crops using these models
based on provided input data.
 Semi-physical Models:
A semi-physical model in remote sensing and GIS is a
type of model that combines physical principles with
remotely sensed data to estimate or predict biophysical
parameters, such as crop yield, biomass. These models are
often used to monitor and manage natural resources, as well
as to assess the impacts of climate change and other
environmental stressors.
 Ensemble Techniques Application:
 Model Averaging: Calculate the simple mean of
predictions from ML, semi- physical model and CSM
models.
 Weighted Averaging: Assign weights based on individual
model performance and calculate the weighted average of
predictions.
A. Semi Physical Approach-NPP:
Fig 3: FAPAR for Latur during Kharif 2023
IJISRT24MAR1981 www.ijisrt.com
International Journal of Innovative Science and Research Technology
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
 Stacking: Use a meta-model that takes predictions from
individual models as inputs and predicts the final yield.
 Voting: Each model votes for a final yield prediction, and
the most frequent prediction is considered.
 Model Validation
 Split the dataset into training, validation, and test sets to
avoid overfitting and ensure generalizability.
 Use metrics like Root Mean Squared Error (RMSE), and
R-squared (R2) for evaluation.
 Assess performance using the test dataset and ground
truth data.
 Quality Control
 Calculate the normalized RMSE between the observed
and ensemble model's estimated yield.
 Ensure RMSE does not exceed acceptable thresholds,
refining the model if necessary.
 Validation
The accuracy of our model was evaluated based on crop
cutting experiment data (CCE data) of PMFBY (Pradhan
Mantri Fasal Bima Yojana) for the crop season kharif-2023.
III. RESULTS AND DISCUSSION
Following were the results and conclusion for different
methods/models used for estimation of yield of soyabean crop
in Latur districts of Maharashtra, Revenue-Circle wise.
2481
Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Fig 4: PAR for Latur during Kharif 2023

Fig 5: Tstress for Latur during Kharif 2023

IJISRT24MAR1981 www.ijisrt.com 2482

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Fig. 6: Waterstress for Latur during Kharif 2023

Fig. 7: Soybean Crop Mask of Latur during

IJISRT24MAR1981 www.ijisrt.com 2483

Volume 9, Issue 3, March – 2024
ISSN No:-2456-2165
Fig. 8: Soybean Yield of Latur during Kharif 2023 kharif 2023
 Ahamadpur, Andhori, and Kingaon demonstrated
relatively higher actual yields, surpassing 2 tonnes per
hectare.
 Hadolati, Belkund, Bhada, Killari, and Shirur Tajband had
comparatively lower actual yields, falling below 2 tonnes
per hectare.
 Semi-Physical (NPP) Yield also showed variation, with
some areas like Ujani, Chakur, and Latur displaying lower
performance.
 Notable outliers include Halgara, which had exceptionally
high actual yields, and Nalgir, which had a notably high
Semi-Physical Yield.
 The dataset suggests a need for closer examination of
factors influencing yield discrepancies between regions,
such as soil quality and agricultural practices.
IJISRT24MAR1981 www.ijisrt.com
International Journal of Innovative Science and Research Technology
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
 Overall, most regions achieved yields above 2 tonnes per
hectare, indicating satisfactory performance for soybean
cultivation in 2023.
 Regional variations in climate and agricultural techniques
likely influenced the observed differences in soybean
yields. Same results were reported by Xiao, X., et.al
(2006) and Yao, Y., et.al (2021)
B. Crop Simulation Model DSSAT-4.8
 The highest yielding RCs are Halgara and Wadhavana
(Bk), both with an average yield of 2.89 tonnes per
hectare.
 The lowest yielding RCs are Nalgir and Tandulja, with
average yields of 0.77 tonnes per hectare and 1.18 tonnes
per hectare, respectively.
2484
Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Fig 9: Soybean Yield in T/ha by DSSAT for Latur during Kharif 2023

 There is a significant variation in yield between different
RCs, with the highest yielding RCs producing more than
four times the yield of the lowest yielding RCs.
 There is a significant variation in yield between different
RCs, with some RCs having double the yield of others.
 The results suggest the potential benefits of using DSSAT
for predicting soybean crop yields, although specific
environmental factors and RC conditions may influence
the accuracy of the predictions. Jadhav, S. D et.al (2018),
Bhosale, A. D., et.al (2015) and Deshmukh, S. D.,
et.al (2013) also elaborated same results for soybean.
C. Machine Learning
 CCE yield and different indices under study showing
accuracy 82 % in Machine learning model. By the method
(SVR) Support Vector Regression accuracy is showing
highest value.

IJISRT24MAR1981 www.ijisrt.com 2485

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Fig 10: Soybean Yield in T/ha by ML for Latur during kharif 2023

 Ahamadpur, Andhori, and Kingaon exhibited relatively
stable yields around 2 tonnes per hectare in both actual
and ML projections.
 Shirur Tajband, Chakur, and Nalegaon showed lower
yields, indicating potential challenges in those regions.
 Halgara stood out with remarkably high actual yield but
substantially lower ML yield, suggesting potential
discrepancies in the ML model for that area.
 Optimization of ML models can enhance predictive
accuracy and contribute to better-informed agricultural
practices.
 The results indicate the potential of ML to improve yield
predictions and optimize crop management strategies.

IJISRT24MAR1981 www.ijisrt.com 2486

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

D. Ensemble Model:

Fig 11: Soybean Yield in T/ha by Ensemble Model for Latur during Kharif 2023

Table 4: Statistical approach give weightage during kharif 2023 as following to different models.
Model Used DSSAT Yield Semi-Physical Yield Machine Learning Yield
Weightages in % 36.72 33.03 30.25

 The Ensemble Yield represents a combination of all above
three predictive models or methods to estimate soybean
crop yield.
 Regions like Tandulja, Hisamabad, and Sakol exhibited
significant negative percent Error, suggesting
considerable discrepancies between actual and predicted
yields.
 Conversely, regions like Walandi and Shelgaon showed
positive percent Error, indicating slight overestimations in
yield predictions.
 Ensemble Yield tended to align closely with Actual yield
in some regions, such as Dewrjan and Borol, where the
percent Error was minimal or zero.
 The summary reveals the potential of ensemble
techniques in predicting soybean yields, though
adjustments may be needed to enhance accuracy in
regions with high % Error.
 Understanding and minimizing % Error can facilitate
better decision-making for farmers and policymakers,
optimizing agricultural practices and resource allocation.
 Continuous refinement and validation of predictive
models can contribute to more reliable yield forecasts,
supporting sustainable soybean production in Latur
district. Same results were given by Md Didarul Islam
et.al (2023), Liujun Xiao et.al. (2022) and Ayan Das a
et.al (2023) in both Machine learning and ensemble
approach.

IJISRT24MAR1981 www.ijisrt.com 2487

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

Table 5: Estimated Yield of Soybean Crop in Tones/Hectors with Different Models and
Percent Error with Ensemble Model for Year 2023
District Tehsil RC Field CCE DSSAT Semi- Machine Ensemble RMSE %
Yield Physical Learning Yield Error
Yield Yield
Latur Ahmadpur Ahamadpur 1.69 2.31 2.22 2.10 2.22 -31
Latur Ahmadpur Andhori 2.14 2.28 2.43 2.06 2.27 -7
Latur Ahmadpur Hadolati 1.77 2.30 1.64 2.10 2.01 -14
Latur Ahmadpur Khandali 1.58 2.66 2.01 2.08 2.28 -44
Latur Ahmadpur Kingaon 2.08 2.33 2.73 2.06 2.40 -15
Latur Ahmadpur Shirur tajband 1.35 2.05 2.01 1.33 1.84 -37
Latur Ausa Ausa 2.12 2.23 2.29 2.15 2.23 -5
Latur Ausa Belkund 1.84 2.18 1.68 2.07 1.97 -7
Latur Ausa Bhada 2.19 2.11 1.87 2.15 2.04 7
Latur Ausa Killari 2.00 2.27 1.73 2.06 2.02 -1
Latur Ausa Kinithot 2.28 2.42 2.14 2.12 2.24 2
Latur Ausa Lamjana 2.18 2.53 1.84 2.12 2.18 0
Latur Ausa Matola 2.08 2.09 1.72 2.08 1.96 6
Latur Ausa Ujani 2.71 2.51 1.70 2.08 2.11 22
Latur Chakur Ashta 1.82 2.51 2.42 2.08 2.36 -30
Latur Chakur Chakur 1.30 2.21 2.68 1.29 2.13 -63
Latur Chakur Nalegaon 1.34 2.60 1.99 1.23 2.02 -51
Latur Chakur Shelgaon 2.26 2.32 2.48 2.12 2.32 -3
Latur Chakur Wadwal (Na) 1.28 2.26 1.97 1.24 1.89 -47
Latur Chakur Zari Bk 1.35 2.74 2.41 1.28 2.24 -66
Latur Deoni Borol 2.04 2.06 2.01 2.15 2.07 -2
Latur Deoni Deoni 1.41 2.33 2.23 1.25 2.01 -42
Latur Deoni Walandi 2.58 2.54 1.98 1.23 1.99 23
Latur Jalkot Ghonasi 2.06 2.19 1.94 2.15 2.09 -1
Latur Jalkot Jalkot 2.37 2.22 2.05 2.11 2.13 10
Latur Latur Babhalgaon 1.57 2.05 1.77 2.06 1.95 -24
Latur Latur Chincholi (Bk) 2.26 2.10 2.10 2.14 2.11 7
Latur Latur Gategaon 1.93 2.26 2.03 2.13 2.15 -11
Latur Latur Harangul (Bk) 1.98 2.10 1.89 2.08 2.02 -2
Latur Latur Kanheri 2.01 2.32 1.70 2.11 2.05 -2
Latur Latur Kasarkheda 1.97 2.57 2.23 2.07 2.32 -18
Latur Latur Latur 2.14 2.14 1.68 2.09 1.97 8
Latur Latur Murud (Bk) 2.20 2.53 1.88 2.10 2.19 1
Latur Latur Tandulja 1.18 2.33 2.36 2.15 2.29 -94
Latur Nilanga Ambulga (Bk) 2.47 2.48 2.47 2.12 2.38 4
Latur Nilanga Aurad (Sha) 2.06 2.06 2.05 2.09 2.07 0
Latur Nilanga Bhutmugli 1.99 2.11 1.87 2.08 2.02 -1
Latur Nilanga Halgara 2.89 3.87 1.91 1.32 2.50 14
Latur Nilanga Kasar Balkunda 2.12 2.59 1.65 2.13 2.13 -1
Latur Nilanga Kasar Shirashi 2.28 2.65 1.92 2.07 2.24 2
Latur Nilanga Madansuri 2.02 2.29 1.75 2.06 2.04 -1
Latur Nilanga Nilanga 2.02 2.23 1.82 2.12 2.05 -2
Latur Nilanga Nitur 2.53 2.61 2.46 2.14 2.43 4
Latur Nilanga Panchincholi 2.39 2.53 2.25 2.14 2.33 3
Latur Renapur Karepur 1.23 2.26 2.66 1.30 2.15 -74
Latur Renapur Palsi 1.39 2.25 2.00 1.93 2.08 -49
Latur Renapur Pangaon 2.18 2.53 2.20 2.09 2.29 -5
Latur Renapur Poharegaon 2.30 2.55 2.06 2.18 2.28 1
Latur Renapur Renapur 1.39 2.65 2.11 1.31 2.11 -52
Latur Shirur-Anantpal Hisamabad 1.09 2.45 2.15 1.29 2.04 -86
Latur Shirur-Anantpal Sakol 1.32 2.66 2.52 1.27 2.24 -70
Latur Shirur-Anantpal Shirur Anantpal 1.53 2.06 2.18 1.29 1.90 -24
Latur Udgir Dewrjan 2.16 2.38 1.93 2.15 2.16 0
Latur Udgir Her 2.39 2.30 2.48 2.15 2.32 3
Latur Udgir Mogha 2.12 2.22 2.03 2.17 2.14 -1
Latur Udgir Nagalgaon 2.40 2.59 2.20 2.15 2.34 2
Latur Udgir Nalgir 1.57 0.77 2.37 2.15 1.70 -8
Latur Udgir Tondar 2.21 2.08 2.33 2.18 2.20 0
Latur Udgir Udgir 2.36 2.19 2.53 2.12 2.29 3

IJISRT24MAR1981 www.ijisrt.com 2488

Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Latur Udgir Wadhavana(Bk) 2.62 2.56
Latur district Average values = 1.968 2.342
Table 6: Average Percent Error of all Approaches Estimated Yield of Soybean Crop for Year 2023.
Methods Field CCE DSSAT Yield Semi-Physical Yield
Yield (T/h) 1.97 2.34 2.11
RMSE % Error -19 -7
In Table 4, the yield estimated by various methods is
presented, including the percentage error of yield by the
ensemble model with field CCE, which is provided in the last [1].
column. Out of 62 points only 14 points were showing more
than 30% error. As per mentioned in deliverables in
YESTECH manual given by Pradhan Mantri Fasal Bima
Yojana, the error (nRMSE) between the observed and
modeled yield should not be more than ±30% for district
level. While, in table no. 5 average yields and % errors for
whole Latur district are mentioned. All values of % error
presenting error was less than ±30%. Which indicates that the [2].
process adopted for RC wise Soybean yield estimation is
acceptable for Latur district for all the models.
IV. CONCLUSION
[3].
"The comparative evaluation of NPP, DSSAT, and
Machine Learning models for predicting soybean crop yields
in Latur, Maharashtra for kharif 2023 has provided a nuanced
understanding of their individual strengths and limitations.
Each model demonstrated distinct capabilities in capturing [4].
the complexities of crop growth dynamics, with Machine
Learning showcasing its adaptability and predictive accuracy.
The ensemble model, combining NPP, DSSAT, and
Machine Learning, offered a holistic perspective by [5].
leveraging the strengths of individual models. This
integration allowed for a more robust and reliable prediction
of crop yields, providing a comprehensive overview of the
crop performance over the study period.
Comparisons between the ensemble model results and
field data revealed a promising alignment, emphasizing the
potential of ensemble modelling in enhancing the accuracy of
yield predictions. The combined approach contributes to [6].
minimizing uncertainties associated with individual models
and provides a more reliable basis for decision-making in
agriculture.
In conclusion, the integration of NPP, DSSAT, and
Machine Learning models into an ensemble framework
presents a promising avenue for advancing crop yield
prediction methodologies. This study serves as a foundation
for further research and refinement, with the ultimate goal of
providing farmers and policymakers with accurate and [7].
actionable insights for sustainable agricultural practices in
Latur District of Maharashtra State."
IJISRT24MAR1981 www.ijisrt.com
https://doi.org/10.38124/ijisrt/IJISRT24MAR1981
2.68 2.12 2.49 5
2.107 1.929 2.149 -14.41
Machine Learning Yield Ensemble Yield
1.93 2.15
2 -9
REFERENCES
Ayan Das, Mukesh Kumar, Amit Kushwaha, Rucha
Dave, Kailash Kamaji Dakhore, Karshan Chaudhari,
& Bimal Kumar Bhattacharya. (2023). Machine
learning model ensemble for predicting sugarcane
yield through synergy of optical and SAR remote
sensing. Remote Sensing Applications: Society and
Environment, Volume 30, April 2023, 100962.
https://doi.org/10.1016/j.rsase.2023.100962
Bhosale, A. D., Waskar, D. P., & Shinde, P. B. (2015).
Performance of DSSAT model for simulating soybean
yield under rainfed condition in Vertisols of central
Maharashtra. International Journal of Agriculture,
Environment and Biotechnology, 8(3), 604-610.
Chavan, K.K., Khobragade, A.M., Kadam, Y.E. and
Mane, R.B. (2018) Study the heat unit requirement of
soybean (Glycine max) varieties under varied weather
condition at Parbhani. Journal of Pharmacognosy and
Phytochemistry. 7(3): 526-530.
Deshmukh, S. D., Waskar, D. P., & Shinde, P. B.
(2013). Application of DSSAT model for soybean
yield prediction in Vertisols of western Maharashtra.
International Journal of Current Microbiology and
Applied Sciences, 2(8), 555-562.
Hoogenboom, G., C.H. Porter, K.J. Boote, V. Shelia,
P.W. Wilkens, U. Singh, J.W. White, S. Asseng, J.I.
Lizaso, L.P. Moreno, W. Pavan, R. Ogoshi, L.A. Hunt,
G.Y. Tsuji, and J.W. Jones. (2019). The DSSAT crop
modeling ecosystem. In: p.173-216 [K.J. Boote,
editor] Advances in Crop Modeling for a Sustainable
Agriculture. Burleigh Dodds Science Publishing,
Cambridge, United Kingdom.
https://dx.doi.org/10.19103/AS.2019.0061.10.
Hoogenboom, G., C.H. Porter, V. Shelia, K.J. Boote,
U. Singh, W. Pavan, F.A.A. Oliveira, L.P. Moreno-
Cadena, T.B. Ferreira, J.W. White, J.I. Lizaso, D.N.L.
Pequeno, B.A. Kimball, P.D. Alderman, K.R. Thorp,
S.V. Cuadra, M.S. Vianna, F.J. Villalobos, W.D.
Batchelor, S. Asseng, M.R. Jones, A. Hopf, H.B. Dias,
L.A. Hunt, and J.W. Jones. (2024). Decision Support
System for Agrotechnology Transfer (DSSAT)
Version 4.8.2 (www.DSSAT.net). DSSAT
Foundation, Gainesville, Florida, USA.
Jadhav, S. D., Waskar, D. P., & Shinde, P. B. (2018).
Evaluation of DSSAT model for soybean yield
prediction under different sowing dates and irrigation
levels in Vidarbha region of Maharashtra. Journal of
AgriSearch, 6(4), 37-42.
2489
Volume 9, Issue 3, March – 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/IJISRT24MAR1981

[8]. Jones, J.W., G. Hoogenboom, C.H. Porter, K.J. Boote,

W.D. Batchelor, L.A. Hunt, P.W. Wilkens, U. Singh,
A.J. Gijsman, and J.T. Ritchie. (2003). The DSSAT
cropping system model. European Journal of
Agronomy, 18, 235-265.
https://doi.org/10.1016/S1161-0301(02)00107-7
[9]. Liujun Xiao, Guocheng Wang, Hangxin Zhou, Xiao
Jin, & Zhongkui Luo. (2022). Coupling agricultural
system models with machine learning to facilitate
regional predictions of management practices and
crop production. Environ. Res. Lett., 17(2022),
114027. https://doi.org/10.1088/1748-9326/ac9c71
[10]. Md Didarul Islam, Liping Di, Faisal Mueen Qamer,
Sravan Shrestha, Liying Guo, Li Lin, Timothy J.
Mayer, & Aparna R. Phalke. (2023). Rapid Rice Yield
Estimation Using Integrated Remote Sensing and
Meteorological Data and Machine Learning. Remote
Sensing, 15(9), 2374. Remote Sensing | Free Full-Text
| Rapid Rice Yield Estimation Using Integrated
Remote Sensing and Meteorological Data and
Machine Learning (mdpi.com)
[11]. Monteith, J. L. (1982). Solar radiation and
productivity in tropical ecosystems. Journal of
Applied Ecology, 19(3), 657-666.
[12]. Nimje, P. M. (2022) Soybean Production technology,
National Skill Development Corporation AISECT,
Agriculture Skill Council of India 2022.
(https://www.nsdcindia.org/scmp/assets/image/19135
97640-
20_SoybeanProductionTechnology_preview.pdf)
[13]. Upasana Singh, Gargi Gaydhane and Ashutosh Pawar
(2023) A Semi - Physical Approach using Remote
Sensing based Net Primary Productivity (NPP),
Spatial, Spectral & Temporal Paddy Yield Model
Development for the State of Assam. International
Journal of Science and Research (IJSR). 12 (8),1175-
1785.
[14]. VNMKV Diary 2023 Krishi Dainandini 2023
published by Vasantrao Naik Marathwada Krishi
Vidyapeeth, Parbhani during 2023.
[15]. Xiao, X., Boles, S., Frolking, S., Li, C., Babu, J. Y.,
Salas, W., & Moore III, B. (2006). Mapping paddy
rice agriculture in South and Southeast Asia using
multi-temporal MODIS images. Remote Sensing of
Environment, 100(1), 95-113.
[16]. Yao, Y., Li, Z., Tian, F., & Tao, F. (2021). Remote
Sensing-Based Estimation of Maize Yield Using a
Semi-Physical Approach: A Case Study in the North
China Plain. Frontiers in Plant Science, 12, 662.

IJISRT24MAR1981 www.ijisrt.com 2490